Sei sulla pagina 1di 14

DETERMINING IMPACTS ON NONPERFORMING LOAN RATIO IN TURKEY

Paylamak Gzeldir :)
Journal of Finance and Investment Analysis, Vol. 2, No.4, 2013
DETERMINING IMPACTS ON NON-PERFORMING LOAN RATIO IN TURKEY
Metin Vatansever1 and Ali Hepen2
Abstract

Banking sector is an essential part of a nations economy and represents


one of the most important components of a nations capital. Similarly, the
loan portfolio represents an important component of a banks total assets.
These assets generate huge interest income which is a critical measure of
the banks financial performance and stability. Therefore, the nonperforming loan ratio is a critical tool to measure a banks performance.
There is recently a growing recognition between macroeconomic
indicators, bank-level factors and the non-performing loans (NPLs) ratio.
The purpose of this study is to investigate whether there is a significant
relationship between macroeconomic indicators, bank-level factors and
non-performing loan ratio in Turkey. In this study linear regression models
and cointegration analysis are utilized to determine the significant
relations between the periods from January 2007 to March 2013. Our
empirical results show that debt ratio, loan to asset ratio, real sector
confidence index, consumer price index, EURO/ Turkish lira rate, USD/
Turkish lira rate, money supply change, interest rate, Turkeys GDP growth,
the Euro Zones GDP growth and volatility of the Standard & Poors 500
stock market index does not have significant effect to explain NPL ratio on
multivariate perspective. On the other hand, industrial production index,
Istanbul Stock Exchange 100 Index, inefficiency ratio of all banks
negatively affect NPL ratio; unemployment rate, return on equity and
capital adequacy ratio positively affect NPL ratio.
JEL classification numbers: C53, E00, E27, E29

Keywords: Turkey, Non-performing loans, Macroeconomic indicators,


Bank specific indicators, Cointegration
1 Introduction
There is a growing recognition that the quantity or percentage of nonperforming loans (NPLs) is related to bank failures and the financial status
of a country. Especially after current global financial crisis, which started in
US and spread to whole world especially Europe, the issue of nonperforming loans (NPLs) has gained increasing attentions because of the
rapid increased default of sub-prime mortgage loans.
Moreover, there are some evidences that financial and banking crisis in
East Asia and Sub Saharan African countries were preceded by increasing
non-performing loans. From the point of this view, the non-performing loan
ratio is therefore a critical measure to evaluate a banks performance, the
economic activity, and the national financial stability and soundness. In
the literature macroeconomic indicators and bank specific factors may
cause increase in NPLs.
From this point and the necessity, the aim of this study is to determine the
long term effects of macroeconomic and bank specific factors on nonperforming loan ratio in Turkey. In particular, we run linear regression
models and cointegration analysis to find a significant and long term
relations between NPL ratio and several specific factors by using time
series dataset covering the monthly period between January 2007 and
March 2013.
The structure of the rest of this research is organized as follows. The next
section provides a literature review that attempts to delineate the
determinants of loan portfolio quality and NPL ratio. Section three
provides overview the data set and theoretical framework adopted in this
paper and section four gives the empirical results. Finally, section five
offers concluding comments.
2 Literature Review
This section reviews the previous empirical studies on determining factors
of the NPLs. Many studies investigate the factors that induce NPLs by
examining potential links between bank-specific variables and
macroeconomic factors.

[1] study finding a regression equation to explain the trends of NPL ratio in
Hang Kong by using nominal interest rates, the CPI, property prices, equity
prices, number of bankruptcies, the unemployment rate, and real GDP as
explanatory variables They find that the NPL ratio rises with increasing
nominal interest rates and an increasing number of bankruptcies, but
decreases with higher CPI inflation, economic growth, and property price
inflation. Additionally, [2] shows that Czech NPL of the corporate sector
rate can be positively affected by increasing real effective exchange rate,
the loan to GDP ratio, unemployment and interest-rate increases.
Beside on classic linear regression model, [3] and [4] use VAR
methodology to investigate which factors is most effect on NPLs. [5]
indicates yields can be used to make accurate predictions of the future
effects of the business cycle on asset quality. [6] show the impact of GDP
growth and the business cycle on credit risk and also on the quality of
bank loans.
On the other hand, [7] and [8] study panel date set to understand of NPLs
behaviors on their own researches. They use both macroeconomic and
bank-specific factors. According to their studies, the quality of loans can
be explained mainly by macroeconomic variables.
3 Data and the Theoretical Framework
3.1 Overview of Data
There is a growing literature which suggests that NPL ratio maybe be
explained by both macroeconomic and bank specific factors. In this study,
we take into consideration, 6 bank specific factors, 10 macroeconomic
factors and 2 global factors. Table 1, table 2 and table 3 gives those
variables.
Table 1: Bank Specific Factors
Bank Level Factors
INEF

Dept

ROE

Definitions

LOAS

CAR

Table 2: Macroeconomic Factors


Macroeconomic Factors

Definitions

RSCI

Confidence Index-Real Sector

CPI

Consumer Price Index

EUR

Euro/ Turkish Lira Rate

USD

USD/ Turkish Lira Rate

IPI

Industrial Production Index

ISE

Istanbul Stock Exchange100 Index

M3Y

Money Supply Change

UR

Unemployment Rate

IR

Interest Rate

GNP

Gross National Product Growth

Table 3: Global Factors


Global Factors

Definitions

EGNP

The Euro Zones GDP Growth

VIX

Volatility of the Standard & Poors 500 Stock Market


Index

Based on countries financial condition and legislation, non-performing


loans term (NPLs) can be different. In Turkey, loan performing loans are
defined as a loan that has been unpaid for ninety days or more. As a
target variable, we use the non-performing aggregate loan ratio
calculated by dividing non-performing loans by total aggregate (including
consumer, housing, auto, credit cards and the other loans) and nonperforming loans. The target and the other factors time period cover the
monthly period January 2007 March 2013, a total of 75 observations and
are collected from the official web site of [9] and [10].
Table 4: Summary Statistics of the NPL Ratio
Series:

NPL

Sample

2007:01 2013:03

Observations:

75

Mean

0.034576

Median

0.033225

Maximum

0.050747

Minimum

0.025157

Std. Dev.

0.007677

Skewness

0.693126

Kurtosis

2.279410

Jarque-Bera

7.627949

Probability

0.022060

Table 4 summaries some descriptive statistics of NPL ratio and presents


that the average ratio, minimum and maximum value is roughly 3.46%,
2.25% and 5.07% respectively, during the period of January 2007 to March
2013. Figure 1 indicates how the NPL ratio is distributed across the time
period.
Figure 1: Total NPL Ratio in Turkey, January 2007-March 2013
As seen from the figure 1, there is one significant jump in the NPL ratio in
Turkey, which occurs in the starting period of the Global Financial Crisis
that originated in the USA with the 2007 collapse of the sub-primes
mortgage market. It shows how the global crises effects NPL ratio in
Turkish credit market. However, NPL ratio goes down slightly after the
global crises. The economy quickly recovered and continued to grow over
the next following months, which is reflected in the improvement of the
NPL ratio.
3.2 Theoretical Framework
3.2.1 Cointegration Analysis
The concept of cointegration and the implications of cointegrating
relationships are very relevant in the real estate market. Real estate
economic and investment theory often suggests that two or more
variables would be expected to hold some long-run relationship with one
another. Therefore, cointegration analysis is a crucial tool for the existence
of such a long-run relationship [11]. There are a number of methods for
testing cointegration in the literature. This article considers two most

commonly used tests of cointegration; namely Engle-Granger (EG) or


Augmented Engle-Granger (AEG) test and Cointegrating Regression
Durbin Watson (CRDW) test [12].
3.2.2 Engle-Granger (EG) or Augmented Engle-Granger (AEG) Test
Regression of a nonstationary time series on other nonstationary time
series may produce a spurious regression. To avoid the spurious
regression problem that may arise from regressing a nonstationary time
series on one or more nonstationary time series, we have to transform
nonstationary time series to make them stationary. If we subject our time
series data individually to unit root analysis and find that they are all I(1);
i.e., they contain a unit root; there is a possibility that our regression can
still be meaningful (i.e., not spurious) provided that the variables are
cointegrated. In order to find out whether they are cointegrated or not, we
simply carry out our original regression and subject our error term to unit
root analysis. If it is stationary; i.e., I(0), it means that our variables are
cointegrated and have a long-term, relationship between them. In short,
provided that the residuals from our regression are I(0) or stationary, the
conventional regression methodology is applicable to data involving
nonstationary time series [13].
3.2.3 Cointegrating Regression Durbin Watson (CRDW) Test
An alternative and quicker method of testing for cointegration is the
CRDW test, whose critical values were first provided by [14]. In CRDW, the
Durbin Watson statistics d obtained from cointegrating regression is
used[1]. But the null hypothesis is d=0 that rather than the standard d=2.
The 1% critical value to test the hypothesis that the true d=0 is 0.511.
Thus, if the computed d value is smaller than 0.511, we reject the null
hypothesis of cointegration at the 1% level. Otherwise, we fail to reject
the null, meaning that the variables in the model are cointegrated and
there is a long-term relationship between the variables [15].
4 Empirical Results
4.1 Development of the Model
The aim of this study is to investigate whether there are significant longterm effects (which is divided into three parts such as country, global and
bank specific factors) on non-performing aggregate loan ratio in Turkish
banking systems. To find the effects between them, NPL ratio is regressed
on those bank specific, macroeconomic and global indicators.

The simplest regression model can be written for NPL ratio as


(1)
where NPLs are non-performing aggregate loan ratio; is the error term
and the subscript t represent time. is intercept term; and finally are the
estimators of confidence index-real sector, consumer price index, Euro/
Turkish Lira rate, USD/ Turkish Lira rate, industrial production index,
Istanbul Stock Exchange 100 Index, money supply %, unemployment rate,
gross national product growth, the euro zones GDP growth, volatility of
the Standard & Poors 500 Stock Market Index (VIX), return on equity,
loan-to-assets ratio, inefficiency, debt and capital adequacy ratio
respectively.
By using Eviews 4.1 statistical package, we run regression analysis on
equation 1. According to the estimated ordinary least square results, pvalues of are respectively all within acceptable range and they are
significance at 5% significance level. On the other hand, the rest of
variables are not significance at 5% significance level. So we ignore those
insignificance variables in those modes. By ignoring them, we obtain the
following estimated ordinary least square results for equation 1.
(2)
At equation 2, P-values of are respectively all within acceptable range and
they are significance at 5% significance level. Finally, p value of the Fstatistics is zero for all the three models. Additionally, we have estimated
more models in order to determine the right specification, by choosing
from the different models estimated R-Squared, Adjusted R-Squared, the
Akaike, Schwarzs Bayesian information criteria. After experimenting with
various functional forms, from R-Squared, Adjusted R-Squared, the Akaike
and Schwarz criteria point of view, the proper model to best adjust the
data below is specified and estimated.
(3)
where is the first lag of error term, and t is a trend that increases by one
for each observation.
By using Eviews 4.1 statistical package, we obtain the following estimated
ordinary least square results for the above equation:

(4)
This last model is obviously the best one. p-values of , , , , , , , , and are
all within acceptable range and they are significance at 5% significance
level. As for goodness-of measures and values are about 0.97 and
0.96 respectively, which indicate the regression fits quite well. Finally, pvalues of the F-statistics are zero. The Akaike, Schwarzs Bayesian
information criteria are -0.901 and -0.597 respectively which are minimum
values in experimenting with various functional forms. It obviously shows
that the first, second and fourth lag of error term have significant and
important effect on our model.
There is, however, a possibility that the ordinary least square results may
be misleading due to inappropriate standard errors because of the
presence of heteroskedasticity. In order to test whether error terms are
heteroskedastic or not the heteroskedasticity test (with cross term) is
carried out [16]. The probability value of 0.209 in this test show that error
term is jointly insignificant even at 5% significance level, meaning that
they are no heterosketastic in our models.
We need also to test for serial correlation. Breusch Godfrey Serial
Correlation LM test is applied for the three different models. The
(effectively) zero probability values in this test strongly indicate the
presence of serial correlation in the residuals. In the presence of serial
correlation, the ordinary least square estimators are still unbiased as well
as consistent and asymptotically normally distributed, but they are no
longer efficient, meaning that standard errors are estimated in the wrong
way and, therefore, usual confidence intervals and hypotheses tests are
unreliable. Moreover, usually, the finding of autocorrelation is also an
indication that the model is misspecified. [17] proposed a general
covariance estimator. The covariance estimator is used to try overcome
standard errors for autocorrelation in the error terms in the model. In
order to correct the standard errors for autocorrelation, the model is reestimated by ordinary least square with Newey-West2 procedure and then
all indicators become significant at 5% significance level. Based on these
results, the model is correct specified.
4.2 Cointegration Analysis
As indicated before, since it is critical to find out whether the results
obtained from our model are meaningful (i.e., not spurious) or not, let we

apply formal unit root tests in each series to test the reliability of our
estimates.
4.2.1 Unit Root Tests
The established standard procedure for cointegration analysis is to start
with unit root tests on the time series data being analyzed. The
Augmented Dickey-Fuller (ADF) and Phillips-Perron (PP) unit root test are
used to test for the presence of unit roots and establish the order of
integration of the variables in the model.
Table 5: Summary of ADF and PP Tests for Unit Roots in the Variables
Variable

ADF Test

Probability

PP Test

Probability

Results

NPL

-0.535

0.481

-0.638

0.437

Fail to reject the null

IPI

1.068

0.924

0.741

0.872

Fail to reject the null

ISE

1.075

0.925

1.091

0.927

Fail to reject the null

UR

-0.195

0.612

-0.474

0.507

Fail to reject the null

ROE

-1.570

0.109

-1.659

0.092

Fail to reject the null

INEF

-0.963

0.297

-0.906

0.321

Fail to reject the null

CAR

-1.279

0.184

-1.330

0.169

Fail to reject the null

Table 5 shows the results of the ADF and PP unit root tests3. The null
hypothesis of the test is that there is a unit root against the alternative
one that there is no unit root in the variables.
The ADF and PP statistics for NPL, IPI, ISE, UR, ROE, INEF and CAR are all
insignificant at 5% level of significance, which leads to non-rejection of the
null hypothesis that there is a unit root problem in the variables.
According to ADF and PP test, it is obvious that the variables are nonstationary.
As mentioned previously, differencing has the effect of making the
variable stationary. Table 6 summarizes the results of unit root tests for
first difference variables.

Table 6: Summary of ADF and PP Tests for Unit Roots in the Variables (In
1st Difference)
Variable

ADF Test

Probability

PP Test

Probability

Results

NPL

-2.176

0.029

-4.047

0.000

Reject the null

IPI

-2.232

0.026

-15.303

0.000

Reject the null

ISE

-3.480

0.000

-7.439

0.000

Reject the null

UR

-.5.825

0.000

-3.178

0.002

Reject the null

ROE

-5.742

0.000

-9.484

0.000

Reject the null

INEF

-4.610

0.000

-4.610

0.000

Reject the null

CAR

-6.157

0.000

-6.185

0.000

Reject the null

The ADF and PP test statistics for the first difference variables are all
significant at 5% level of significance, which leads to rejection of the null
hypothesis that there is a unit root problem in the variables. Based on ADF
and PP test, it is apparent that the first difference variables are stationary,
which implies that the variables are integrated of order one, I(1).
4.2.2 The Augmented Engle-Granger (AEG) Test
The residuals from the estimation of equation 3 used to test for the
existence of cointegrating relationship between the NPLs ratio and several
macroeconomic and bank specific factors. The null hypothesis is that the
residuals have a unit root problem against the alternative that the
variables cointegrate. The AEG test is presented in the below table (Table
7).
Table 7: Summary of ADF and PP test output for 3 equation
ADF Test

Probability

PP Test

Probability

Results

-2.770

0.006

-8.138

0.000

Reject the null

The probability values in those tests indicate that residuals are significant
at 5% significance level, meaning that that the null hypothesis is rejected.
To reject the null hypothesis implies that the residuals have not a unit root
problem, i.e., they are stationary. It can therefore be concluded that,
based on the AEG method, the variables are cointegrated.
4.2.3 Cointegrating Regression DurbinWatson Test
Since cointegration is very crucial to the reliability of estimated
parameters, a second test, namely CRDW test, was carried out to make
sure that the variables in this study are definitely cointegrated. The
DurbinWatson statistic for the regression represented by equation (3) is
1.61, which is above the 1% critical value of 0.511. Therefore, we fail to
reject the null hypothesis of cointegration at the 1% level.

To sum up, our conclusion is based on both the AEG and CRDW tests
giving the variables that NPL, IPI, ISE, UR, ROE, INEF and CAR are
cointegrated. Depending on these results, we may infer that the
appropriate model for NPL is the one represented in equation (3) and
determines that our estimations are reliable, i.e., not spurious.
Equation 4 reflects that unemployment rate, return on equity, inefficiency,
capital adequacy rate have positive long-term effect on NPL ratio. In
figures, when unemployment rate (UR), return on equity (ROE), capital
adequacy (CAR) increased by 1 point then NPL rate increased 0.15, 0.011
and 0.146 by point respectively and when industrial production index (IPI),
Istanbul Stock Exchange100 Index (ISE), Inefficiency ratio of all banks
(INEF) increased by 1 point then NPL ratio decreased by 0.004, 0.109 and
0.063 point, respectively
In addition, equation 4 reveals that there is a significant positive relation
between NPL rate and the first, second and also fourth lag of error term. In
figures, when , and increased by 1 point, the DRPPI increased by 1.118,
0.522 and 0.272 point respectively.
5 Conclusion
This study analyzes the relationship between the NPLs ratio and several
macroeconomic and bank specific factors in Turkey by using ordinary least
square estimation approach with integration analysis and the time series
from January 2007 to April 2013.
Empirical results show that that debt ratio, loan to asset ratio, confidence
index-real sector, consumer price index, EURO/ Turkish lira rate, USD/
Turkish lira rate, money supply change, interest rate, GDP growth, the
Euro Zones GDP growth and volatility of the Standard & Poors 500 stock
market index does not have significant effect to explain NPL ratio on
multivariate perspective.
On the other hand, industrial production index (IPI), Istanbul Stock
Exchange 100 Index (ISE), Inefficiency ratio of all banks (INEF) negatively,
Unemployment rate (UR), return on equity (ROE), capital adequacy ratio
(CAR) positively affect NPL ratio.
Additionally the positive and negative effects are such a long-term, not
spurious. Our findings have several implications in terms of policy and
regulation. It can help identify the causes of NPL ratio and thus lead

analysts, policymakers, investors and financial institutions to a better


understanding of banking and credit market conditions as well as their
impact on economic activity, and the national financial stability and
soundness.
References
[1] S. Gerlach, W. Peng and C. Shu, Macroeconomic conditions and
banking performance in Hong Kong SAR: A panel data study,
BIS Papers, 22, Bank for International Settlements, Basel, (2005).
[2] P. Jakubik, Macroeconomic environment and credit risk, Czech Journal
of Economics and Finance,10(1), (2007), 133166.
[3] M. Gambera, Simple forecasts of bank loan quality in the business
cycle, Emerging Issue Series, 3, Federal Reserve Bank of Chicago,
Chicago, (2000).
[4] W. Blaschke and M. Jones, Stress testing of financial systems: An
overview of issues, methodologies and FSAP experiences, IMF Working
Paper, 01/88, Washington, (2001).
[5] M. Gambera, Simple forecasts of bank loan quality in the business
cycle, Emerging Issue Series, 3, Federal Reserve Bank of Chicago,
Chicago, (2000).
[6] W. Blaschke and M. Jones, Stress testing of financial systems: An
overview of issues, methodologies and FSAP experiences, IMF Working
Paper, 01/88, Washington, (2001).
[7] D.P. Louzis, A.T. Vouldis and V.L. Metaxas, Macroeconomic and bankspecific determinants of non-performing loans in Greece: A comparative
study of mortgage, business and consumer loan portfolios, Journal of
Banking & Finance, 36(4), (2011), 1012-1027.
http://dx.doi.org/10.1016/j.jbankfin.2011.10.012
[8] N. Klein, Non-performing loans in CESEE: Determinants and impact on
macroeconomic performance, IMF Working Paper, 13/72, Washington,
2013.
[9] Banking Regulation and Supervision Agency (BRSA), [Online]
Available from: http://www.bddk.org.tr/WebSitesi/English.aspx (October
21, 2013).

[10] Central Bank of the Republic of Turkey (CBRT), [Online] Available


from: http://www.tcmb.gov.tr/yeni/eng/ (October 25, 2013).
[11] C. Brooks and S. Tsolacos, Real estate modeling and forecasting, First
edition, Cambridge University Press, New York, 2010.
[12] D.N. Gujurati, Basic econometrics, Fourth edition, The McGrawHill,
New York, 2004.
[13] D.N. Gujurati, Basic econometrics, Fourth edition, The McGrawHill,
New York, 2004.
[14] J.D. Sargan and A. Bhargava, Testing residuals from least squares
regression for being generated by the gausian random
walk, Econometrica, 51(1), (1983), 153-174.
http://dx.doi.org/10.2307/1912252
[15] D.N. Gujurati, Basic econometrics, Fourth edition, The McGrawHill,
New York, 2004.
[16] H. White, A heteroscedasticity-consistent covairance matrix estimator
and a direct test for heteroscedasticity, Econometrica, 48(4), (1980), 817838. http://dx.doi.org/10.2307/1912934
[17] W.K. Newey and D.K. West, A simple, positive semi-definite,
heteroskedasticity and autocorrelation consistent covariance
matrix, Econometrica, 55(3), (1987), 703708.
http://dx.doi.org/10.2307/1913610

Yildiz Technical University, Faculty of Arts and Science, Department of


Math and Stats. Istanbul. e-mail: vatansevermetin@gmail.com
2 Assoc.Prof.Dr., Istanbul University, School of Business, Department of
Finance. Istanbul. e-mail: alihepsen@yahoo.com
[1] We know that , so if there is to be a unit root, the estimated is about
1, which implies that d is about zero.
1

2 Newey and West procedure may not appropriate in small samples. Since
we have 75 observations, our samples may be regarded as reasonable
large.
3 Two lags have been used in ADF unit root tests.